|Publication number||US6148395 A|
|Application number||US 08/856,645|
|Publication date||Nov 14, 2000|
|Filing date||May 15, 1997|
|Priority date||May 17, 1996|
|Publication number||08856645, 856645, US 6148395 A, US 6148395A, US-A-6148395, US6148395 A, US6148395A|
|Inventors||Tuan Q. Dao, Donald E. Steiss|
|Original Assignee||Texas Instruments Incorporated|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (4), Referenced by (28), Classifications (11), Legal Events (4)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims priority under 35 USC 119(e) (1) of the provisional application Ser. No. 60/017,459, filed May 17, 1996.
This invention is in the field of microprocessors, and is more specifically directed to single-chip multiprocessor integrated circuits.
In the field of data processing systems and circuits, parallel processing refers to the simultaneous use of multiple data processing circuits, such as microprocessor central processing units (CPUs), to perform a single computer program. In this arrangement, the multiple CPUs execute program instructions in parallel, under certain constraints relative to the order of operations and use of shared resources. The theoretical performance of a parallel processing system thus can be much higher than that of systems having a single CPU. Supercomputers having many parallel processors (e.g., up to on the order of 256 or more processors) are known in the art.
Many of the known parallel processing systems are arranged using multiple microprocessor integrated circuits, with each integrated circuit having a single CPU. It is contemplated, however, that future parallel processing systems will be implemented, at least in part, by single-chip microprocessor integrated circuits that have multiple CPUs implemented therewithin.
The benefits of parallel processing have been achieved, in part, in modern single chip microprocessors of the so-called "superscalar" architecture. For example, microprocessors of the well-known x86-architecture, particularly those compatible with and having the capability of the PENTIUM microprocessor available from Intel Corporation, are considered to be superscalar. Superscalar microprocessors include at least two instruction pipelines that can operate in parallel, under certain constraints upon the types of parallel instructions (considering the available processor resources). A two-pipeline superscalar microprocessor is thus able to execute up to two multiple-cycle instructions in each machine cycle.
Typical modern microprocessors also include on-chip floating-point units (FPUs) that perform arithmetic operations on floating-point data operands. As is fundamental in the art, floating-point data operands are digital data words that have both a mantissa and an exponent portion, for representing non-integer numbers. Typically, as in the case of 486-class microprocessors, the FPU is implemented in a microprocessor as a separate execution unit, having its own instruction pipeline, such that floating-point instructions are forwarded to the FPU for decoding and execution, in similar fashion as if it were located on a separate integrated circuit. In the case of the PENTIUM microprocessor, which is superscalar in the sense that it has two integer pipelines, the floating-point pipeline is shared with one of the integer pipelines through the integer execution and integer writeback stages, with the floating-point execution stages being added to the length of that integer pipeline.
As is well known in the art, the FPU of a modern microprocessor is a relatively complex circuit, and requires a significant portion of the integrated circuit chip area for its realization. According to conventional techniques, therefore, the implementation of multiple CPUs, each having their own dedicated FPU, onto a single integrated circuit chip, will require an extremely large chip size. This large chip size translates, of course, into both low manufacturing yields and also high per-chip manufacturing cost.
The invention may be implemented into a single-chip microprocessor integrated circuit having multiple integer central processing units (CPUs) sharing a floating-point unit (FPU). The shared FPU provides a register file, such as a stack, corresponding to each of the CPUs; in addition, the shared FPU has input instruction buffers for receiving instructions from each of the CPUs. The shared FPU has one execution pipeline that performs floating-point operations upon operands received from any one of the multiple CPUs, according to instructions received from the associated CPU. Dispatch circuitry is provided to detect and resolve conflicts among the multiple CPUs in their use of the shared FPU.
It is therefore an object of the present invention to provide a parallel processing microprocessor architecture that allows for a floating-point unit to be shared among multiple integer central processing units on the same integrated circuit.
It is a further object of the present invention to reduce the integrated circuit chip area required to implement a high performance parallel processing microprocessor, by sharing the floating-point unit.
It is a further object of the present invention to obtain the benefit of parallel processing with reduced chip area, considering the duty cycle with which parallel integer processors use the floating-point unit.
Other objects and advantages of the present invention will be apparent to those of ordinary skill in the art having reference to the following specification together with its drawings.
FIG. 1 is an electrical diagram, in block form, of a single-chip multiprocessor according to a first preferred embodiment of the invention.
FIG. 2 is a flow diagram illustrating the sharing of pipeline stages by the floating-point unit in the first preferred embodiment of the invention.
FIG. 3 is an electrical diagram, in block form, of a shared floating-point unit as incorporated into the multiprocessor of FIG. 1.
FIG. 4 is a timing diagram illustrating an example of the progression of floating-point instructions through the floating-point unit of FIG. 3.
FIG. 5 is an electrical diagram, in block form, of a single-chip multiprocessor according to a second preferred embodiment of the invention.
FIG. 6 is an electrical diagram, in block form, of a shared floating-point unit as incorporated into the multiprocessor of FIG. 5.
Referring first to FIG. 1, single-chip multiprocessor 2 according to the preferred embodiment of the invention will now be described in detail. For purposes of this description, the term "single-chip multiprocessor" is intended to refer to a single integrated circuit chip upon which two or more microprocessor central processing units, or CPUs, are implemented. In this example, multiprocessor 2 includes two x86 architecture superscalar CPUs, 100, 101, each capable of executing sequences of program instructions independently from one another (subject to memory and other resource constraints, as are known in the art). As will be evident from the following description, the present invention may be applicable to single-chip multiprocessor arrangements including more than two CPUs, and including CPUs of architecture types other than the well-known x86 architecture.
Multiprocessor 2 is in communication with an external bus XBUS by way of bus interface unit (BIU) 8. External bus XBUS is connected to system memory, disk storage devices, input/output devices, and other conventional subsystems typical in conventional personal computers and workstations. External bus XBUS thus is capable of communicating address, data, and control signals between multiprocessor 2 and such other subsystems in the conventional manner. Of course, other buses external to multiprocessor 2 may also be provided, examples of which include the well-known PCI local bus. BIU 8 is connected to internal bus IBUS, which is a bus for communicating address, data, and control signals among the various constituent blocks of multiprocessor 2, including CPUs 100, 101. Multiprocessor 2 according to this embodiment of the invention includes shared floating-point unit (FPU) 5, the construction of which will be described in further detail hereinbelow, and cache memory 6, each of which are connected to internal bus IBUS. Cache 6 is an on-chip cache memory, either having dedicated data and instruction portions or instead being a unified cache, for locally storing data and instructions that are either recently used or likely to be used in the particular instruction sequences being executed by multiprocessor 2; data and address information will be communicated to and from cache 6 on internal bus IBUS, as shown in FIG. 1.
General control and timing logic is also included within multiprocessor 2, but is not shown in FIG. 1 for purposes of clarity. Such general control and timing logic will include logic for generating the clock signals within multiprocessor 2, responsive to which synchronous operation may be performed. Such logic will also maintain general control of multiprocessor 2, including arbitration for traffic upon internal bus IBUS and for the use of certain shared resources, and will update status information regarding the current state of the device.
Each CPU 10 includes various functional circuitry of the x86 architecture type to perform a sequence of data processing instructions. Load/store units 12 are included within CPU 10 to control memory access by its associated CPU 10. In addition, each CPU 10 may include its own cache memory 13, for storing either or both data and instructions for the current operation. In this exemplary embodiment, instruction fetch circuitry 15 and instruction decoder 16 support multiple integer instruction pipelines. Instruction decoder 16, in each CPU 10, decodes the instruction fetched by fetch circuitry 15 from cache 13 or elsewhere into control signals communicated to and applied by control unit 14 to control the operation of CPU 10.
As noted above, each CPU 10 is of the superscalar type, and as such is capable of issuing and executing multiple instructions in parallel, absent resource conflicts or data dependencies. In this example, each x86-architecture CPU 10 includes two integer execution units 17, 18, each of which include circuitry for performing operand fetch, execution, and writeback operations by way of one or more ALUs, and also by way of load/store units 12. As such, each CPU 10 effectively has two integer instruction pipelines. Each CPU 10 includes its own register file 20, which includes the usual named x86 architecture registers for retaining data, control, and status information. According to this exemplary implementation, therefore, up to two instructions per machine cycle may be executed by each CPU 10.
It is of course contemplated that the preferred embodiment of the invention is applicable to CPUs 10 having different levels of parallel pipelines, including scalar CPUs (i.e., a single pipeline per CPU), and superscalar CPUs of greater degree (e.g., microprocessors having up to as many as four parallel execution pipelines). It is therefore to be understood that the dual pipeline superscalar arrangement of CPUs 10 in this embodiment of the invention is presented merely by way of example.
It has been discovered, in connection with the present invention, that the on-chip floating-point unit of a conventional single-CPU microprocessor is utilized only for a relatively fraction of the time. In other words, the high-performance FPU in a modern microprocessor, for example of the PENTIUM class, is idle much of the time. According to the present invention, multiple CPUs on the same integrated circuit chip should therefore be able to share a single high-performance FPU without significant degradation in overall multiprocessor performance, while saving the chip area that would otherwise be required to provide a dedicated FPU, even of modest capability, for each CPU in the multiprocessor. The incorporation of a shared FPU into a single-chip multiprocessor according to the present invention is therefore advantageous, especially considering that the control circuitry required to implement the sharing is less than the area that would be required to implement dedicated a per-CPU floating-point unit.
Referring back to FIG. 1, this sharing of a single FPU 5 by multiple CPUs 10 according to a first preferred embodiment of the invention will now be described. In this embodiment of the invention, execution units 170, 171 of CPUs 100, 101, respectively, are in communication with shared FPU 5, while execution units 180, 181 are not. In effect, FPU 5 shares portions of the integer pipelines corresponding to execution units 170, 171 in both CPUs 100, 101, similarly as does the floating-point unit in a single CPU microprocessor of the PENTIUM type. Referring now to FIG. 2, this sharing is illustrated relative to a pipeline diagram.
As shown in FIG. 2, shared FPU 5 operates according to instruction pipeline 40 that shares its initial stages with superscalar integer pipelines 300, 301 of CPUs 100, 101. Superscalar integer pipeline 300 (corresponding to CPU 100) will now be described in detail, with the understanding that pipeline 301 (corresponding to CPU 101) operates similarly. The first stage of pipeline 300 is instruction fetch stage 320, in which instruction fetch circuitry 150 fetches the next instruction to be decoded and executed (by each of the integer pipelines) from memory. The instruction fetch of stage 320 is preferably made from one of on-chip cache memories 6, 130, if the next instruction is available therein; otherwise, the instruction is fetched from main memory via BIU 8 and external bus XBUS. Predecode stage 340 is next performed by instruction decoder 160 to at least partially decode the instructions fetched in stage 320. The decoded instructions are then sent to the pair of integer pipelines corresponding to execution units 170, 180 for further decoding and processing.
As noted above, according to this example, shared FPU 5 operates in conjunction with only one of the two execution pipelines in superscalar pipeline 300 of CPU 100, but not with the other. In the example of FIG. 2, therefore, execution unit 180 thus is available only to execute non-floating-point operations. In the integer pipeline corresponding to execution unit 180, decode second stage 36a0 is first performed to further decode the instruction, retrieve the appropriate operands, and to issue the instruction to execution stage 38a0. Execution stage 38a0 operates in conjunction with one or more integer arithmetic logic units to perform the arithmetic or logical operation specified by the current instruction; the results of execution stage 38a0 are then written back to the appropriate register in register file 200 or to memory, as the case may be, in writeback stage 39a0. The integer pipeline corresponding to execution unit 170, including decode second stage 36b0, execute stage 38b0, and writeback stage 39b0, similarly operates upon non-floating-point instructions detected by predecode stage 340.
Accordingly, the two integer execution pipelines of CPU 100 share initial pipeline stages (e.g., instruction fetch stage 320 and predecode stage 340), but have separate stages for the remainder of their operation. As such, two integer-based instructions may be performed substantially in parallel by execution units 170, 180 of CPU 100, subject to certain instruction pairing constraints, such as data dependencies and resource conflicts.
As noted above, CPU 101 operates according to superscalar pipeline 301, in similar fashion as superscalar pipeline 300 of CPU 100.
Floating-point pipeline 40 of FIG. 2 corresponds to the general sequence of operation of shared FPU 5 in cooperation with CPUs 100, 101. The first stage of floating-point pipeline 40 is performed simultaneously for floating-point instructions detected by instruction decoders 160, 161, in their respective integer predecode stages 340, 341. As such, instruction queue stage 410 receives a series of instruction codes for floating-point instructions detected in predecode stage 340, while instruction queue stage 411 receives a series of instruction codes for floating-point instructions detected in predecode stage 341. It is preferred that multiple floating-point instructions are queued in stages 41; for example, in the case of floating-point pipeline 40 of FIG. 2 where the maximum instruction latency within the pipeline is four instructions, the preferred depth of instructions in queue stages 41 is five instructions. Each sequence of floating-point instructions are then decoded, in FIFO order, in decode stages 420, 421, of floating-point pipeline 40; as indicated in FIG. 2, these stages may be performed simultaneously and in parallel for instructions from each CPU 10.
The next floating-point pipeline stage is dispatch stage 44. Dispatch stage 44 is performed by control circuitry in CPU 5, as will be described hereinbelow, to detect potential conflicts arising from floating-point instructions provided by CPUs 10, and to select and issue the appropriate instruction to the execution stages of the pipeline. Dispatch stage 44 will also effect arbitration between instruction sequences from CPUs 100, 101 in the event of simultaneous floating-point requests. Upon selection of a floating-point instruction sequence and at such time as dependencies are clear, dispatch stage 44 of floating-point pipeline 40 receives operands from the writeback pipeline stage 39b associated with the one of CPUs 100, 101 whose floating-point instruction is to be issued by dispatch stage 44. The communication of operands are illustrated in FIG. 2 by flow arrows OP0, OP1 from pipelines 300, 301, respectively.
It should be noted that floating-point pipeline stages 41, 42, 44 may be, and preferably are, performed in parallel with predecode, decode, and execute stages 34b, 36b, 38b in superscalar integer pipelines 30.
The selected floating-point instruction is then issued by dispatch stage 44 to the execution stages of floating-point pipeline 40. In this example, operand preprocessing stage 45 is first performed, to format the operands received in dispatch stage 44 (and also operands in the internal registers or on the writeback bus of shared FPU 5 itself) for application to the execution circuitry. Special case operands (such as NaN, infinity, exact zero, and denormalized values) are also identified and handled by operand preprocess stage 45. In this example, three execution stages 46, 47, 48 are included within floating-point pipeline 40, indicating that the floating-point arithmetic instructions may require up to three cycles to execute; of course, single-cycle instructions (such as change sign) may also be performed, in which case certain of execution stages 46, 47, 48 may be skipped. Completion stage 49 is the final stage of floating-point pipeline 40, in which the results of the instruction are written back to the appropriate destination location, and in which any execution exceptions or errors are identified and communicated to the appropriate resource.
Referring now to FIG. 3, the construction of shared FPU 5 of multiprocessor 2, which in this embodiment of the invention is operable according to the pipeline flow of FIG. 2, will now be described in detail.
As shown in FIG. 3, shared FPU 5 includes two multi-stage instruction buffers 500, 501 that receive instruction codes from instruction decoders 160, 161 of CPUs 100, 101, respectively. As noted above relative to the description of queue stages 41 in floating-point pipeline 40, instruction buffers 50 are preferably arranged in a FIFO manner, having a number of stages equal to the maximum instruction latency of shared FPU 5, plus one. In this example, each of instruction buffers 50 have five stages, and thus can hold five instruction codes, given that the maximum instruction latency of shared FPU is four instructions (operand preprocess stage 45 plus the three execution stages 46, 47, 48). Each stage of instruction buffers 50 is preferably able to store an entire instruction code for a floating-point instruction which, in this case of x86-architecture CPUs 10, may include up to eight bytes (one for the tag, and up to seven bytes for the instruction code with identifiers for source and destination operand locations).
Shared CPU 5 also includes floating-point instruction decoders 520, 521 associated with a corresponding instruction buffer 500, 501. Floating-point instruction decoders 52 each receive the oldest instruction from their respective instruction buffers 50, and decode it into control signals for application within shared FPU 5; the control signals are applied, simultaneously from both instruction decoders 520, 521 to dispatch unit 56 in shared FPU 5. Floating-point instruction decoders 50 also control the receipt of input operand data from internal bus IBUS via input buffer 54. In this example, input buffer 54 is preferably an eighty-bit register for storing a floating-point data word presented thereto via internal bus IBUS. However, since internal bus IBUS is only sixty-four bits wide in this example, multiple loads of data from internal bus IBUS into input buffer 54 are required. Floating-point instruction decoders 56 are each capable of controlling input buffer 54 to receive additional data from internal bus IBUS if necessary to fill input buffer 54; such control is effected by each of instruction decoders 52 by way of lines FMR0, FMR1 shown in FIG. 3. Accordingly, fetching of extended precision floating-point operands by shared FPU 5 is effected by one of floating-point instruction decoders 52 controlling input buffer 54 to receive a first portion of the operand in a first cycle and, via asserting a signal on the corresponding FMR line, fetch the remainder of the extended precision operand in the next cycle.
As evident from the foregoing description, shared FPU 5 includes both an instruction buffer 50 and a floating-point instruction decoder 52 for each of the CPUs 10 that share FPU 5. As such, receipt and queuing of floating-point instructions and decoding of these instructions may be performed by shared FPU 5 in parallel, as described hereinabove relative to FIG. 2. In this embodiment of the invention, dispatch unit 56, which receives control signals from each of floating-point instruction decoders 52, determines which of the floating-point instructions decoded by decoders 52 shall next be issued to the execution circuitry of FPU 5, as will now be described. It is contemplated that one of ordinary skill in the art, having reference to this description, will be able to readily implement the logic required for dispatch unit 56 based upon the functions now to be described therefor.
Dispatch unit 56 is constructed of logic suitable for performing the functions of receiving control signals from floating-point instruction decoders 52 that are derived from floating-point instructions, and for launching, or issuing, the appropriate instruction to the execution circuitry of shared FPU 5, along with the appropriate operands; as such, dispatch unit 56 performs the functions described hereinabove for dispatch stage 44 of shared floating-point pipeline 40. Dispatch unit 56 also receives control signals from instruction buffers 50, upon which are communicated signals indicating the full or empty state of instruction buffers 50, for use by dispatch unit in a priority determination. Based upon the control signals received from instruction decoders 52 and from instruction buffers 50, dispatch unit 56 first determines if floating-point instructions have been decoded by more than one of instruction decoders 52. If so, dispatch unit 56 will select which of the multiple instructions (i.e., which of CPUs 10) has the higher priority, according to a predetermined arbitration scheme such as round-robin, fixed order, or instruction-dependent.
For either a single floating-point instruction or for the selected one of the multiple floating-point instructions, dispatch unit 56 next determines if any dependencies are presented by the selected instruction relative to those instructions that are in execution. If so, the floating-point pipeline is stalled until completion of an earlier instruction, or other action is taken in order to avoid the problem of the dependency. Dispatch unit 56 also interrogates the instruction to determine if it is a microcoded instruction, in which case dispatch unit 56 will forward the appropriate identifier to microcode sequencer 57 which, in turn, will present the appropriate control signals corresponding to the microcoded instruction to dispatch unit 56.
Dispatch unit 56 is also responsible for selecting the appropriate sources for the two operands (OPA, OPB) to be applied to the execution circuitry. In this embodiment of the invention, multiplexer 58a selects the appropriate source for operand OPA and multiplexer 58b selects the appropriate source for operand OPB, each according to control signals generated by dispatch unit 56. Each of multiplexers 58a, 58b receive data retrieved from memory and communicated on internal bus IBUS as an eighty-bit input from input buffer 54 for use as operands OPA, OPB, respectively. Constant ROM 59 is provided in shared FPU 5 to store floating-point representations of commonly used numbers such as π, e, √2, and the like; the ROM address for the desired constant stored in constant ROM is presented by dispatch unit 56 responsive to the decoded selected floating-point instruction. Also according to this embodiment of the invention, the results of previously executed instructions are presented on writeback bus WB and applied to the inputs of multiplexers 58 for potential use as floating-point operands, so as to provide operand forwarding without requiring storage of the result in an additional register.
According to this embodiment of the invention, shared FPU 5 includes dedicated register banks for each CPU 10. In this example, the register banks are implemented as stacks 600, 601, dedicated for CPUs 100, 101, respectively, considering that floating-point registers in conventional x86-architecture microprocessors are implemented as stacks. In this example, stacks 600, 601 each include eight register locations, each capable of storing a floating-point data word (in the maximum available precision), and also the necessary support registers such as a stack pointer and the like. Each of stacks 600, 601 have outputs for presenting a floating-point data word to each of multiplexers 58a, 58b for use as input operands OPA, OPB, depending upon the floating-point instruction to be launched. Since the floating-point instruction itself will not refer to a selected one of stacks 60, but since the proper stack 60 will be implicit in the one of CPUs 10 from which the current floating-point instruction was forwarded, dispatch unit 56 must maintain CPU identity for the instruction to be launched.
Of course, the dedicated register banks may be implemented in other ways, such as by way of multiple named registers, depending upon the particular FPU architecture and instruction set.
Dispatch unit 56 launches an instruction by controlling multiplexers 58 to select the appropriate operands for use as operands OPA, OPB in the execution of the floating-point instruction, and by sending the appropriate control signals for that instruction to control stage 0 of control pipeline 66. Control pipeline 66 is a pipeline of control circuitry that produces the appropriate control signals to be applied to the execution circuitry in FPU 5 for the instructions in the pipeline. Referring back to FIG. 2, since four execution stages 45, 46, 47, 48 are used in shared FPU 5 according to this embodiment of the invention, control pipeline 66 includes four stages, as shown in FIG. 3.
The execution of the launched instruction is initiated by operand preprocessor 62, which reformats operands OPA, OPB as necessary. For example, shared FPU 5 may utilize an internal format of higher complexity than extended precision (or whatever is the highest precision level externally from FPU 5), so as to maintain improved resolution for intermediate results of a multipass floating-point operation. If such is the case, operand preprocessor 62 receives operands OPA, OPB from multiplexers 58 and reformats them into the internal format (operands OPA', OPB'). In addition, operand preprocessor 62 also examines operands OPA, OPB for special cases, such as not-a-number (NaN), infinity, exact zero, and the like, and presents an corresponding special value as operand OPA' or OPB' at its outputs. This operation of operand preprocessor 62, which occurs in 0th execution stage 45 (FIG. 2), is controlled by control signals from the corresponding stage of control pipeline 66.
Operands OPA', OPB' are then presented to execution circuitry 65 of shared FPU 5 so that the appropriate arithmetic or logic operation may be performed. Execution circuitry 65 of FPU 5 may be implemented in any conventional manner for performing floating-point operations. In this example, execution circuitry 65 is implemented in three paths, each of which receives operands OPA', OPB' from operand preprocessor 62. The one of these multiple paths which is to perform the operation is selected and enabled by the corresponding stages of control pipeline 66.
In this example, a first path of execution circuitry 65 is multiplication circuitry 70, for performing floating-point multiplication and division operations. Multiplication circuitry 70 may include a sequence of circuitry known in the art, such as a Booth recorder, multiplier arrays, an adder, and rounding circuitry. In this particular example, multiplication circuitry 70 provides a result in three machine cycles, and thus operates according to the three execution stages 46, 47, 48 of floating-point pipeline 40, each stage under the control of a corresponding stage in control pipeline 66. The output of the rounding circuitry in multiplication circuitry 70 is presented to completion unit 75.
A second path of execution circuitry 65 is implemented as adder circuitry 72, for performing additive and subtractive operations. In this example, adder circuitry 72 includes a sequence of an aligner, an adder, a LEO (left end out) shifter, normalization circuitry and rounding circuitry; of course, as in the case of multiplication circuitry 70, other conventional implementations of floating-point additive circuitry may be utilized. Also as in the case of multiplication circuitry 70, the execution of an instruction by adder circuitry 72 may take three machine cycles, and as such adder circuitry 70 may be implemented as three stages, each controlled by a corresponding stage in control pipeline 66.
The third path of execution circuitry 65 is a single-cycle execution unit 68, by way of which special single-cycle floating-point instructions may be executed. Examples of such single-cycle floating-point instructions include floating-point change of sign (FCHS) and floating-point absolute value (FABS). Single-cycle execution unit 68 is implemented by way of the special combinational logic required to perform the desired single-cycle instructions, as known in the art. The output of single-cycle execution unit 68 is also presented to completion unit 75.
Completion unit 75, as noted above, receives the outputs of the three paths (multiplication circuitry 70, adder circuitry 72, and single-cycle execution unit 68) of execution circuitry 65. Completion unit 75 includes such logic as necessary to detect and handle any errors or exceptions encountered in the execution of the instruction. In addition, if shared FPU 5 is capable of out-of-order execution, completion unit 75 will include the necessary logic and local storage to re-order the execution results into program order. Each of these operations are performed in completion stage 49 of floating-point pipeline 40 (see FIG. 2). Completion unit 75 also may reformat the results of the execution, as required, into the appropriate format. The results from completion unit 75 are output to result register 76 which, in turn, drives writeback bus WB at its output.
As shown in FIG. 3, writeback bus WB is coupled to output buffer 78, through which communication of the result of the floating-point operation to memory, via internal bus IBUS, may be effected. Writeback bus WB is also coupled to each of dedicated stacks 600, 601, for register storage of the results. As noted above, shared FPU 5 also provides result forwarding, and as such writeback bus WB is coupled to inputs of multiplexers 58 as a source of input operands for later instructions.
Referring now to FIG. 4, the operation of shared FPU 5 in executing instructions provided from each of CPUs 10 that share it as a resource will now be described relative to a pipeline diagram for an example of a program sequence. Each row of FIG. 4 corresponds to a machine cycle, while each column of FIG. 4 is a pipeline stage in floating-point pipeline 40 of FIG. 2. In the example of FIG. 4, each of CPUs 100, 101 provides shared FPU 5 with a sequence of floating-point instructions identified as FPIk(0) for instructions from CPU 100 and as FPIk(1) for instructions from CPU 101 ; the index k indicates the position of the instruction in the sequence, with the first instruction having an index k of 1.
In cycle n of this example, the final stage of instruction queue pipeline stage 410 for instructions from CPU 100 contains, in instruction buffer 501, the first instruction FPI1(0). The remainder of the floating-point pipeline stages are empty at this time (presumably because shared FPU 5 has not been utilized for several machine cycles). The final stage of instruction queue pipeline stage 411 for instructions from CPU 101 is also empty in cycle n, as no floating-point instruction forwarded from CPU 101 has yet advanced to this point. Accordingly, in the next cycle n+1, the first instruction FPI1(0) is decoded by floating-point instruction decoder 521 in floating-point instruction decode pipeline stage 42. In this example, a first floating-point instruction FPI1(1) from CPU 101 has advanced to the final stage of instruction queue pipeline stage 411 ; similarly, the second floating-point instruction FPI2(0) from CPU 100 has advanced into the final stage of instruction queue pipeline stage 410.
In cycle n+2, the first floating-point instruction FPI1(0) is launched by dispatch unit 56 in dispatch stage 44. In this example, the sequence of instructions from each of CPUs 100, 101 continue to advance in the pipeline, with new instructions FPI3(0), FPI2(1) advancing into the final queue stages 410, 411, respectively. As noted above, dispatch stage 44 determines the priority between floating-point instructions decoded by the multiple instruction decoders 52 of shared FPU 5. In the example of FIG. 4, no conflict was present because only instruction FPI1(0) was decoded by instruction decoder 521 and presented to dispatch unit 56 in cycle n+1, and thus instruction FPI1(0) can be issued in cycle n+2 (assuming no exceptions or other faults).
Execution of instruction FPI1(0) thus begins in cycle n+3, in stage 45. In the immediately previous cycle n+2, however, instruction decoders 520, 521 have decoded floating-point instructions FPI2(0), FPI1(1), respectively. Accordingly, a conflict is present at dispatch stage 44 in cycle n+3. In this example, dispatch unit 56 resolves the conflict by selecting the second instruction FPI2(0) from CPU 100 for processing in dispatch stage 44 in cycle n+3, as it is the next instruction in the sequence of the instruction FPI1(0) that was launched in the previous cycle. Accordingly, the sequence of instructions from CPU 101 stalls, and instruction FPI1(1) remains in decode stage 42 until the sequence of instructions from CPU 100 is either complete or reaches a point at which instructions from CPU 101 may be processed by shared FPU 5.
In the example of FIG. 4, the sequence of floating-point instructions from CPU 100 is three instructions long. In cycle n+3, therefore, instruction FPU3(0) has reached decode stage 420, but the final queue stage 410 is empty. According to this embodiment of the invention, therefore, the sequence of instructions FPI1(0) through FPI3(0) is processed in sequence until cycle n+5, at which time no conflict remains at dispatch stage 44. In cycle n+5, the first instruction in the sequence from CPU 101 is launched, and the sequence of instructions being executed by shared FPU 5 thus begins to include the instructions from CPU 101. As shown in FIG. 4, therefore, particularly in cycles n+7 and n+8, shared FPU 5 is able to include, in its pipeline, instructions from multiple CPUs 10.
According to the preferred embodiment of the invention, therefore, multiple floating-point sequences may be executed by shared FPU 5. It is contemplated, based on analysis done in connection with the present invention, that the degree of sharing (i.e., the number of CPUs sharing a single FPU) will be selected by way of a tradeoff between chip area saved and performance loss due to conflicts. Given the frequency of FPU use by conventional processors, it is contemplated that 2:1 or 3:1 sharing ratios will not result in frequent conflicts. Even in the event of a conflict, however, it is also contemplated that the wait time required of the CPUs may be minimized by allowing portions of floating-point sequences to be interleaved. Given the tremendous savings in chip area that may be obtained in the sharing of floating-point units, however, it is expected that the present invention will be extremely beneficial in the implementation of single-chip multiprocessors.
Referring now to FIGS. 5 and 6, the implementation of multiprocessor 102 according to a second embodiment of the invention will now be described. Like elements in FIGS. 5 and 6 to those in FIGS. 1 and 3 will be referred to by way of the same reference numerals.
Multiprocessor 102 is a single-chip multiprocessor having two CPUs 1100, 1101 implemented therewithin; as described above relative to FIG. 2, more than two CPUs may be implemented in multiprocessor 102 according to this embodiment of the invention, if desired. The degree of resource sharing exhibited by CPUs 110 in multiprocessor 102 is much less than that in multiprocessor 2 described hereinabove. As illustrated in FIG. 5, CPU 1100 is in communication with internal bus IBUS0, which is dedicated to CPU 1100 and which is coupled to external bus XBUS via dedicated BIU 1080. Similarly, CPU 1101 is in communication with its own dedicated internal bus IBUS1, which is coupled to external bus XBUS via dedicated BIU 1081. Each of CPUs 1100. 1101 is also coupled to dedicated cache memory 1060, 1061, respectively.
BIUs 1080, 1081 each operate as a full bus interface unit for a microprocessor, and as such control and effect the communication of data, address, and control signals between internal buses IBUS0, IBUS1, respectively, and external bus XBUS. BIUs 108 each access external bus XBUS as separate bus masters.
According to this embodiment of the invention, CPUs 110 are illustrated as having similar construction as CPUs 10 described hereinabove. As described hereinabove, however, the internal construction of CPUs 10 may be according to other known arrangements for microprocessor CPUs. For example, CPUs 110 may be scalar microprocessors, or superscalar microprocessors to a higher degree than two (e.g., having four integer pipelines). Other alternative implementations for CPUs 110 may thus be utilized in connection with the present invention.
According to this second embodiment of the invention, shared FPU 105 not only receives instruction streams from CPUs 110 via buses IS0, IS1, as before, but also communicates data with CPUs 1100, 1101 via internal buses IBUS0, IBUS1, respectively. Referring now to FIG. 6, the construction of shared FPU 105 according to this embodiment of the invention will now be described in detail.
Shared FPU 105 according to this embodiment of the invention is similarly constructed relative to its dispatch unit 56 and execution circuitry 65 as described hereinabove. Also as described above, dedicated instruction buffers 52, floating-point instruction decoders 52, and stacks 60 are provided within shared FPU 105 for each CPU 110 having access thereto. As such, the pipelined operation of shared FPU 105 according to this second embodiment of the invention is quite similar to that described hereinabove relative to FIGS. 2 and 4. According to this embodiment of the invention, however, additional circuitry is provided to communicate with both of internal buses IBUS0, IBUS1.
Internal bus IBUS0 is coupled to input buffer 1540, which receives floating-point operands therefrom. Floating-point instruction decoder 520, which decodes floating-point instructions presented by CPU 1100, controls input buffer 1540 to enable the receipt of extended precision floating-point operands (having, for example, eighty bits) from the smaller sixty-four bit internal bus IBUS0. The output of input buffer 1540 is presented to inputs of both of multiplexers 58a, 58b to communicate floating-point operands retrieved from memory by CPU 1100. Also on this side of shared FPU 105, output buffer 1780 receives the contents of writeback bus WB and presents its contents to internal bus IBUS0, for communication back to CPU 1100. As indicated by FIG. 5, no communication to CPU 1101 can be effected by internal bus IBUS0.
Shared FPU 105 is in communication with CPU 1101 via internal bus IBUS1, as shown in FIG. 5. In symmetric fashion relative to internal bus IBUS0, internal bus IBUS1 is received by input buffer 1541, which presents its output to inputs of each of multiplexers 58a, 58b; floating-point instruction decoder 521 controls multiple-cycle receipt of floating-point operands by way of a control signal on line FMR1. On the output side, output buffer 1781 is connected to writeback buffer WB to receive results therefrom, and to communicate these results to CPU 1101 via internal bus IBUS1 for receipt by CPU 1101.
Comparison of this second embodiment of FIGS. 5 and 6 with the first embodiment of FIGS. 1 and 3 shows that the present invention is applicable to single-chip multiprocessors of various architectures regarding the extent to which the multiple CPUs share resources such as bus interface units, caches, internal buses, and the like. As repeatedly noted hereinabove, the particular architecture of the sharing CPUs in the multiprocessor can be quite varied, and may be of the CISC (e.g., x86 architecture) or RISC type. In addition, the way in which instructions are decoded and issued by the shared FPU may also be implemented in alternative ways; for example, the integer pipelines may decode the floating-point instructions into so-called atomic operations (in the RISC sense) and the atomic operations may then be forwarded to the shared FPU for issuing to the execution circuitry.
According to the present invention, therefore, important advantages in the design and implementation of single-chip multiprocessors may be obtained. The relatively low utilization of conventional on-chip floating-point units is used to advantage in the present invention, by the sharing of the on-chip floating-point unit by multiple integer CPUs in the multiprocessor. Significant chip area is thus saved according to this arrangement, especially when compared against the alternative of providing a floating-point unit for each integer CPU.
While the present invention has been described according to its preferred embodiments, it is of course contemplated that modifications of, and alternatives to, these embodiments, such modifications and alternatives obtaining the advantages and benefits of this invention, will be apparent to those of ordinary skill in the art having reference to this specification and its drawings. It is contemplated that such modifications and alternatives are within the scope of this invention as subsequently claimed herein.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5598546 *||Aug 31, 1994||Jan 28, 1997||Exponential Technology, Inc.||Dual-architecture super-scalar pipeline|
|US5673407 *||Mar 8, 1994||Sep 30, 1997||Texas Instruments Incorporated||Data processor having capability to perform both floating point operations and memory access in response to a single instruction|
|US5732209 *||May 14, 1996||Mar 24, 1998||Exponential Technology, Inc.||Self-testing multi-processor die with internal compare points|
|US5828578 *||Nov 29, 1995||Oct 27, 1998||S3 Incorporated||Microprocessor with a large cache shared by redundant CPUs for increasing manufacturing yield|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US6223196 *||May 8, 1998||Apr 24, 2001||International Business Machines Corporation||Shared mac (multiply accumulate) system and method|
|US6581155 *||Aug 25, 1999||Jun 17, 2003||National Semiconductor Corporation||Pipelined, superscalar floating point unit having out-of-order execution capability and processor employing the same|
|US6609168 *||Mar 31, 2000||Aug 19, 2003||Intel Corporation||Bus master read request adaptive prefetch prediction method and apparatus|
|US6725354 *||Jun 15, 2000||Apr 20, 2004||International Business Machines Corporation||Shared execution unit in a dual core processor|
|US6728790 *||Oct 15, 2001||Apr 27, 2004||Advanced Micro Devices, Inc.||Tagging and arbitration mechanism in an input/output node of a computer system|
|US6735687 *||Jun 15, 2000||May 11, 2004||Hewlett-Packard Development Company, L.P.||Multithreaded microprocessor with asymmetrical central processing units|
|US6754805 *||Aug 7, 2000||Jun 22, 2004||Transwitch Corporation||Method and apparatus for configurable multi-cell digital signal processing employing global parallel configuration|
|US6907518||Jun 16, 2003||Jun 14, 2005||National Semiconductor Corporation||Pipelined, superscalar floating point unit having out-of-order execution capability and processor employing the same|
|US7117389 *||Sep 18, 2003||Oct 3, 2006||International Business Machines Corporation||Multiple processor core device having shareable functional units for self-repairing capability|
|US7124224 *||Dec 22, 2000||Oct 17, 2006||Intel Corporation||Method and apparatus for shared resource management in a multiprocessing system|
|US7316021 *||Feb 17, 2004||Jan 1, 2008||Sun Microsystems, Inc.||Switching method in a multi-threaded processor|
|US7353364 *||Jun 30, 2004||Apr 1, 2008||Sun Microsystems, Inc.||Apparatus and method for sharing a functional unit execution resource among a plurality of functional units|
|US7353368 *||Feb 15, 2000||Apr 1, 2008||Intel Corporation||Method and apparatus for achieving architectural correctness in a multi-mode processor providing floating-point support|
|US7464208 *||Jan 11, 2006||Dec 9, 2008||Intel Corporation||Method and apparatus for shared resource management in a multiprocessing system|
|US8732368||Feb 17, 2005||May 20, 2014||Hewlett-Packard Development Company, L.P.||Control system for resource selection between or among conjoined-cores|
|US9003168 *||Feb 17, 2005||Apr 7, 2015||Hewlett-Packard Development Company, L. P.||Control system for resource selection between or among conjoined-cores|
|US9152427 *||Oct 15, 2009||Oct 6, 2015||Hyperion Core, Inc.||Instruction issue to array of arithmetic cells coupled to load/store cells with associated registers as extended register file|
|US9715392 *||Aug 29, 2014||Jul 25, 2017||Qualcomm Incorporated||Multiple clustered very long instruction word processing core|
|US20020116438 *||Dec 22, 2000||Aug 22, 2002||Steven Tu||Method and apparatus for shared resource management in a multiprocessing system|
|US20040162971 *||Feb 17, 2004||Aug 19, 2004||Sun Microsystems, Inc.||Switching method in a multi-threaded processor|
|US20040225840 *||May 9, 2003||Nov 11, 2004||O'connor Dennis M.||Apparatus and method to provide multithreaded computer processing|
|US20050066079 *||Sep 18, 2003||Mar 24, 2005||International Business Machines Corporation||Multiple processor core device having shareable functional units for self-repairing capability|
|US20060136925 *||Jan 11, 2006||Jun 22, 2006||Steven Tu||Method and apparatus for shared resource management in a multiprocessing system|
|US20080222336 *||Jan 14, 2008||Sep 11, 2008||Yoshikazu Kiyoshige||Data processing system|
|US20120216012 *||Oct 15, 2009||Aug 23, 2012||Hyperion Core, Inc.||Sequential processor comprising an alu array|
|US20160062770 *||Aug 29, 2014||Mar 3, 2016||Qualcomm Incorporated||Multiple clustered very long instruction word processing core|
|EP1241567A2 *||Jul 19, 2001||Sep 18, 2002||Samsung Electronics Co., Ltd.||Method and apparatus for executing coprocessor instructions|
|EP1241567A3 *||Jul 19, 2001||Sep 10, 2003||Samsung Electronics Co., Ltd.||Method and apparatus for executing coprocessor instructions|
|U.S. Classification||712/222, 712/220, 712/23, 712/221, 712/E09.069, 712/E09.053|
|Cooperative Classification||G06F9/3885, G06F9/3851|
|European Classification||G06F9/38S, G06F9/38E4|
|May 15, 1997||AS||Assignment|
Owner name: TEXAS INSTRUMENTS INCORPORATED, TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:DAO, TUAN Q.;STEISS, DONALD E.;REEL/FRAME:008564/0817
Effective date: 19960513
|Mar 29, 2004||FPAY||Fee payment|
Year of fee payment: 4
|Apr 17, 2008||FPAY||Fee payment|
Year of fee payment: 8
|Apr 24, 2012||FPAY||Fee payment|
Year of fee payment: 12